TracePermitAgentic Visibility & Control

A Java proof of concept that attributes and evaluates AI-agent tool actions before safe, simulated execution.

🛡️

Pre-execution mediation

Supported simulator actions cross authentication, normalization, intent and risk classification, policy, and permit checks before dispatch.

👤

Identity with provenance

Separate a verified human or workload subject from the agent actor and the evidence available at each origin boundary.

📜

Request-bound decisions

Bind decisions and approvals to the exact normalized action; reject changed requests and reused permits.

🔍

Endpoint attestation

The Java stdio connector observes local process facts and signs a request-bound proof in the qualified CLI profile.

🌐

Hosted-client path

One ChatGPT + Auth0 + private-tunnel composition was qualified with OAuth attribution and simulated allow/block actions. Other clients remain to be qualified.

🧪

Safe evidence

Stateful AWS and Jira simulators, bounded evidence, and a read-only React console support repeatable demos without real target credentials.

The enforcement boundary

TracePermit Architecture Overview

The model proposes a tool call; TracePermit governs the call at the controlled MCP boundary. It does not see a hosted model's private reasoning or all of its network traffic. The current POC has no real AWS or Jira executor and is not a production gateway.

Runtime visibility and evidence

TracePermit projects evaluated actions, cryptographic evidence, causal lineage graphs, and deterministic policy decisions into a live read-only console:

TracePermit Console Preview

Start with the reproducible local demo

Use the source quickstart for Java 21, Node.js 22, pnpm, and the numbered simulator scenarios. The deployment status page distinguishes the local POC from a future production topology.

TracePermit is released under the Apache License 2.0.